Artificial intelligence (AI) drove a lot of real data center investments last year, continuing a multi-year trend that has been led by the market’s largest hyperscalers and is forecast to continue surging over at the next several year.

Dell’Oro Group in a new report found that global data center capex increased 51% in 2024, hitting $455 billion in investment last year. The research firm noted that custom accelerators from hyperscalers like Amazon Web Services (AWS), Microsoft, and Google Cloud accounted for more than half of that total investment, and most of last year’s growth was fed by those hyperscalers bolstering their support for AI training workloads.

“While Nvidia’s Hopper architecture – followed by Blackwell systems later in the year – dominated spending, custom accelerators from Google, Amazon, and Microsoft also fueled growth,” Dell’Oro Group’s Baron Fung wrote.

Fung also noted that those hyperscalers increased their investment with colocation providers into infrastructure to support “dedicated AI networks and high-power facilities to support these compute-intensive workloads.”

Dell’Oro Group’s insight echoed that of Synergy Research Group (SRG), which reported that the number of data centers operated by hyperscalers has doubled over the past five years, hitting a precise 1,136 locations at the end of 2024. More importantly, the capacity of hyperscale data centers has doubled over just the past four years, with a more succinct AI-powered uptick over the past several quarters.

“[One hundred and thirty-seven] new hyperscale data centers came online in 2024, continuing a steady trend in growth that goes back many years,” SRG Chief Analyst John Dinsdale wrote. “The big difference now is the increased scale of many of those new data centers. Historically the average size of new data centers was increasing gradually, but this trend has become supercharged in the last few quarters as companies build out AI-oriented infrastructure.”

That scope and scale are also driving energy demands.

Real estate consulting firm JLL noted in a recent report that North America data center tenants gobbled up 4.4 gigawatts (GW) of power last year, which was quadruple of what was consumed in 2020. That demand was met by more than 2.6 GW of colocation capacity that was constructed last year, a “record-setting 6.6 GW of colocation capacity” that under construction at year end, and a pipeline of planned projects that totaled 22.9 GW of additional capacity.

However, meeting those energy demands could hobble future growth plans.

Matt Landek, who leads JLL’s Data Center Project Development and Services, described North America’s power grid as “taxed,” and that power remains the biggest hurdle toward expanding data center capacities.

“We’ve been talking about the power problem for years, but there was still available capacity on the grid,” Landek told SDxCentral. “We’re just slowly working ourselves toward maxing out the grid.”

Landek described ongoing logistical and technical challenges of coordinating the delivery of power resources to new data center locations that can delay the construction of those data centers for years. That has forced many to turn to alternative power sources to act as “bridge solutions,” which can help meet the “speed game” that has emerged.

“I think what we have seen is organizations that want to get connected first,” Landek said. “There’s been a lot of focus in the industry of just getting connected to power.”

Data center market details SRG, for its part, found that the United States remains home to more than half of all worldwide data center capacity as measured by megawatts of “critical IT loads.” China is home to 16% of that capacity, just ahead of Europe.

Dell’Oro Group pointed to Dell Technologies as the leading OEM server provider to data centers based on revenues in 2024, ahead of rivals Hewlett Packard Enterprise (HPE) and Supermicro. The firm also noted that accelerated servers accounted for just over one-third of OEM server revenues, which it tied to growing AI adoption outside of the hyperscale space, and white-box server vendors snared 56% of total server revenues last year.

More data center expansion to come, with a caveat Analyst firms expect data center demand and resulting expansion to continue.

SRG predicts hyperscale data center capacity will double over the next four years, with each year witnessing “a reasonable steady 130-140 additional hyperscale data centers coming online.”

Dell’Oro Group is also bullish on data center spend, forecasting capex will surge more than 30% this year, driven by a “sustained demand on AI infrastructure and a broader recovery in general-purpose infrastructure for servers and networking.” However, it did add that specific enterprise IT spend on data centers might be impacted by ongoing economic uncertainty, supply constraints, and a greater push toward consumption-based services that “could result in near-term fluctuations in AI investments within enterprise data centers.”

That growth could also hit an availability wall.

JLL’s data showed that data center vacancies stood at a record low of just 2.6% at the end of last year, with rents surging 12% year over year.

“Tenants renewing five-year leases are experiencing significant sticker shock, facing up to 50% rent increases, and landlord concessions are becoming increasingly rare in this tight market,” the firm’s report notes.

JLL’s Landek added that this dynamic will impact how enterprises work through their cloud migration strategies.

“Landlords really do hold leverage at this point, and whether that’s a hyperscaler that can take down gigawatts of power or organizations that have deep pockets, they do hold all the cards,” Landek said. “With 70% [of capacity] pre-leased, there’s not a lot of places to go and that’s where the enterprise decision is of what they’re going to do with their compute load, … they’re really at an inflection point of what to do because it’s not going away anytime soon. When you look at the capacity that came on in [2024], near 100% was leased by the time it came online, and that was mostly – from a wholesale perspective – with the hyperscalers. So, it’s a legitimate problem.”